from __future__ import absolute_import
import numpy as nm
import scipy.sparse as sp
from scipy.sparse.linalg import aslinearoperator
from scipy.linalg import eigvals_banded
from sfepy.base.base import get_default, output
from sfepy.linalg import infinity_norm
from six.moves import range
[docs]
def sym_tri_eigen(diags, select_indices=None):
"""
Compute eigenvalues of a symmetric tridiagonal matrix using
`scipy.linalg.eigvals_banded()`.
"""
if select_indices is not None:
n = diags.shape[1]
select_indices = nm.minimum(select_indices, n)
eigs = eigvals_banded(diags, lower=True, select='i',
select_range=select_indices)
else:
eigs = eigvals_banded(diags, lower=True, select='a')
return eigs
[docs]
def cg_eigs(mtx, rhs=None, precond=None, i_max=None, eps_r=1e-10,
shift=None, select_indices=None, verbose=False, report_step=10):
r"""
Make several iterations of the conjugate gradients and estimate so
the eigenvalues of a (sparse SPD) matrix (Lanczos algorithm).
Parameters
----------
mtx : spmatrix or array
The sparse matrix :math:`A`.
precond : spmatrix or array, optional
The preconditioner matrix. Any object that can be multiplied by
vector can be passed.
i_max : int
The maximum number of the Lanczos algorithm iterations.
eps_r : float
The relative stopping tolerance.
shift : float, optional
Eigenvalue shift for non-SPD matrices. If negative, the shift is
computed as :math:`|shift| ||A||_{\infty}`.
select_indices : (min, max), optional
If given, computed only the eigenvalues with indices `min <= i <= max`.
verbose : bool
Verbosity control.
report_step : int
If `verbose` is True, report in every `report_step`-th step.
Returns
-------
vec : array
The approximate solution to the linear system.
n_it : int
The number of CG iterations used.
norm_rs : array
Convergence history of residual norms.
eigs : array
The approximate eigenvalues sorted in ascending order.
"""
n_row = mtx.shape[0]
norm = nm.linalg.norm
rhs = get_default(rhs, nm.random.rand(n_row))
i_max = get_default(i_max, min(n_row, 100))
if shift is not None:
if shift < 0.0:
mtx_norm = infinity_norm(mtx)
output('matrix max norm:', mtx_norm, verbose=verbose)
shift = abs(shift) * mtx_norm
output('eigenvalue shift:', shift, verbose=verbose)
mtx = mtx + shift * sp.eye(n_row, n_row, dtype=mtx.dtype)
mtx = aslinearoperator(mtx)
lambda_max = 0
lambda_min = 0
econd = 1.0
x0 = nm.zeros_like(rhs)
r = r0 = rhs
# The diagonals (0, 1) in two rows.
diags = nm.empty((2, i_max + 1), dtype=mtx.dtype)
diags[0, 0] = 0
if precond is None:
z0 = r0
else:
z0 = precond * r0
x = x0
z = z0
p = nm.zeros_like(rhs)
beta = 0.0
rho0 = nm.dot(z0, r0)
norm_r0 = norm(r0);
if verbose:
output('%5d lambda: %e %e cond: %e |R|: %e\n' % (0, 0, 0, 0, norm_r0))
norm_rs = [norm_r0]
for ii in range(i_max):
p = z + beta * p
q = mtx * p
alpha = rho0 / nm.dot(p, q)
if (not nm.isfinite(alpha)) or abs(alpha) < 1e-16:
output('precision limit reached!')
ii -= 1
break
x = x + alpha * p
r = r - alpha * q
if precond is None:
z = r
else:
z = precond * r
norm_r = norm(r)
norm_rs.append(norm_r)
rho1 = nm.dot(z, r)
beta = rho1 / rho0
# Lanczos
diags[0, ii] += 1.0 / alpha
diags[1, ii] = - nm.sqrt(beta) / alpha
diags[0, ii+1] = beta / alpha
if verbose and (ii > 0):
if (ii % report_step) == 0:
eigs = sym_tri_eigen(diags[:, :ii+1],
select_indices=select_indices)
if select_indices is None:
lambda_min, lambda_max = eigs[0], eigs[-1]
econd = lambda_max / lambda_min
output('%5d lambda: %e %e cond: %e |R|: %e\n'
% (ii, lambda_min, lambda_max, econd, norm_r))
else:
output('%5d |R|: %e\n'
% (ii, norm_r))
if (norm_r / norm_r0) < eps_r:
output('converged on %d iters, |Ri|/|R0|: %e, econd: %e\n'
% (ii, norm_r / norm_r0, econd), verbose=verbose)
break
rho0 = rho1
eigs = sym_tri_eigen(diags[:, :ii+1], select_indices=select_indices)
if verbose and (select_indices is None):
lambda_min, lambda_max = eigs[0], eigs[-1]
econd = lambda_max / lambda_min
output('min: %e max: %e cond: %e\n'
% (lambda_min, lambda_max, econd))
if shift is not None:
eigs -= shift
return x, ii, nm.array(norm_rs), eigs